Paper_Pricing Central Tendency in Volatility


Published on

NES 20th Anniversary Conference, Dec 13-16, 2012
Article "Pricing Central Tendency in Volatility" presented by Stanislav Khrapov at the NES 20th Anniversary Conference.
Author: Stanislav Khrapov, New Economic School

  • Be the first to comment

  • Be the first to like this

No Downloads
Total Views
On Slideshare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Paper_Pricing Central Tendency in Volatility

  1. 1. Pricing Central Tendency in Volatility∗ Stanislav Khrapov† New Economic School September 17, 2012 Abstract It is widely accepted that there is a risk of fluctuating volatility. There is some evi- dence, analogously to long-term consumption risk literature or central tendency in inter- est rates, that there exists a slowly varying component in volatility. Volatility literature concentrates on investigation of two-factor volatility process, with one factor being very persistent. I propose a different parametrization of volatility process that includes this persistent component as a stochastic mean level (central tendency). The reparametriza- tion is observationally equivalent but has compelling economic interpretation. I estimate the historical and risk-neutral parameters of the model jointly using GMM with the data on realized volatility and VIX volatility index and treating central tendency as completely unobservable. The main empirical result of the paper is that on average the volatility pre- mium is mainly due to the premium on highly persistent shocks of the central tendency. Keywords: stochastic volatility; central tendency; volatility risk premium; GMM JEL Classification: G12 ∗ We are grateful to seminar participants at Midwest Econometrics Group (Chicago, October 2011), SoFiEAnnual Meeting (Oxford, June 2012), ICEF/LFE Workshop (Moscow, July 2012), EEA-ESEM (Malaga, August2012) for helpful comments, discussions and suggestions. † Address: New Economic School, Nakhimovskiy Prospekt 47, Moscow, Russia, 117418.Phone: +7 (495) 956 9508. Email:
  2. 2. 1 IntroductionIt is widely accepted that there is a risk of fluctuating volatility. There is some evidence, analo-gously to long-term consumption risk literature (Bansal and Yaron, 2004) or central tendency ininterest rates (Balduzzi et al., 1998)1 , that there exists a slowly varying component in volatility.Bollerslev and Mikkelsen (1996) and Andersen and Bollerslev (1997) find strong evidence of highvolatility persistence. Volatility literature concentrates on investigation of two-factor volatilityprocess, with one factor being very persistent. I propose a different parametrization of volatil-ity process that includes this persistent component as a stochastic drift or central tendency ofmarket volatility. The reparametrization is likely observationally equivalent but has compellingeconomic interpretation. With this model I am able to price two volatility components sepa-rately.2 I estimate the historical and risk-neutral parameters of the model jointly using GMMwith the data on realized volatility and VIX volatility index and treating central tendency ascompletely unobservable. The main result of the paper is that on average the volatility premium is at large extentcomes from the premium on highly persistent shocks of the central tendency. Additional shortlived but very volatile shocks bear a small but statistically significant premium. The volatilitypremium in most part compensates for the shocks in stochastic volatility drift rather than shocksof fast mean reversion to this central tendency. Hence, the role of shock persistence is crucialin determining the compensation for volatility risks. The model I propose is very similar in structure to Duffie et al. (2000) and Bollerslev et al.(2010)3 who are generally concerned with memory patterns in stock market volatility and volatil-ity premium. They employ continuous-time general equilibrium approach together with Epsteinand Zin (1989) time non-separable preferences. These preferences are a crucial feature of themodel that allows to separate volatility and volatility of volatility risk premia. In my modelthe second priced factor is central tendency, or stochastic drift of market volatility. Althoughthe stochastic discount factor in my model does not come from the particular assumption oninvestor preference structure, it implies a similar compensation structure for different sourcesof volatility risk. Analogously to continuous time interest rate model of Cox et al. (1985) it became widespread in financial literature to model stochastic volatility as a mean-reverting process aroundconstant mean level. The seminal work in this direction is done by Heston (1993) who proposedstochastic volatility continuous-time option pricing model. It is also well known that market 1 See also Andersen and Lund (1997); Reschreiter (2010, 2011). 2 Adrian and Rosenberg (2008) find some evidence that volatility components risk, both short and long-run,are priced by analyzing a cross-section of portfolio returns. 3 For analogous discrete time approach see Tauchen (2005) and Bollerslev et al. (2009). 2
  3. 3. volatility is highly persistent and has a thick tailed stationary distribution. Moreover, it iswidely accepted that one-factor stochastic volatility models do not fit well and can not capturehigh persistence and thick tails at the same time. The idea of multi-factor volatility modeldates back to Engle and Lee (1996) who consider several specifications of continuous stochasticvolatility model. One of the specifications includes two additive volatility factors one of thembeing very persistent. The models are discretized using Euler approximation to match GARCHform and estimated with QML. I propose a continuous-time model of market volatility where drift is not constant but ratherstochastic and is driven by a separate mean-reverting stochastic process with its own randominnovation. Andersen and Lund (1997) develop a continuous-time model of interest rates thathas a stochastic drift instead of constant mean level inside of the interest rate dynamics. Inthis paper I do the same for stochastic volatility of the return. The appealing interpretation ofsuch modeling approach, in contrast to additive component representation, is the interpretationof central tendency as a stochastic mean of volatility which determines the average level ofvolatility for a prolonged period of time. Engle and Lee (1999) propose GARCH-like specification of stochastic volatility with uncon-ditional mean replaced with slowly varying second GARCH component. In this specification thedifference between two components is interpreted as transitory volatility component. One of thekey limitations of this model is that only one innovation term drives both volatility componentswhich does not play well with an idea of several sources of volatility risk. In a similar modelingapproach Christoffersen et al. (2008) stress the result that a two-component model fits betterthan a one-component model with jumps. The disadvantage of GARCH models is that they are not closed under temporal aggregation(Drost and Nijman, 1993) and parameter estimates are critically dependent on sampling interval.In this paper I derive exact discretization of the model with stochastic drift. Discretized jointmodel of volatility and central tendency is a vector autoregression of the order one with movingaverage heteroscedastic error structure of order one. The error structure is also kept in explicitform of stochastic integrals. Gallant et al. (1999) estimate two-factor additive stochastic volatility model using EfficientMethod of Moments and find that this model may successfully account for long memory ef-fect. Chernov et al. (2003) evaluate empirically several continuous-time model specifications ofstochastic volatility. In particular, some specifications include two additive volatility factors.One factor is responsible for tail thickness of returns, the other reflects volatility persistence.Corsi (2009) propose an additive cascade model of several volatility components that have dif-ferent effect depending on the time horizon. He shows that despite the absence of genuine longmemory the model is very successful in reproducing empirical characteristics of the returns. 3
  4. 4. Duffie et al. (2000) in conclusion to the paper propose a two-factor model of stochastic volatil-ity model where one factor plays a role of stochastic trend rather than just an additional additivefactor. They argue that given sufficiently small speed of mean reversion this factor may capturelong memory in volatility which is argued to be evident in the data. Bates (2000) employ thecontinuous-time model of S&P500 index options with two-factor additive stochastic volatility.Volatility risk premium is assumed to be exogenous and proportional to current volatility. Thisassumption is consistent with simple log utility. Model parameters are estimated implicitlythrough minimization of option pricing errors. When estimating stochastic volatility models the question of measurement is critical. Clearly,point-in-time volatility in continuous-time model is unobservable. Instead one has to use someapproximations or implied measures. Andersen and Bollerslev (1998) give theoretical justifica-tion for approximation of integrated volatility using high-frequency return data. Bollerslev andZhou (2002)4 propose an elegant approach to estimate parameters of structural continuous-timemodel of returns with stochastic volatility. The main idea is to express moment conditions interms of integrated volatility rather than point-in-time values. Historical integrated volatility ismeasured as realized volatility or standard deviation of high frequency returns over daily period.Jiang and Oomen (2007) approach the problem of latent variable estimation from a differentperspective but also on the basis of high-frequency data and GMM. Joint estimation of the model parameters requires not only historical observation of marketvolatility but also a risk-neutral expectation of volatility. Risk-neutral measure relative to ob-jective measure provides a link to investor preference parameters. Britten-Jones and Neuberger(2000) provide the theoretical justification for the model-free measure of integrated volatilitywhich only requires current option prices. Carr and Wu (2009) generalize this approach anduse it to analyze historical dynamics of variance risk premia of multiple indexes and individualstocks. The general idea of model-free measurement is to use a large set of option prices toconstruct a volatility measure. This measure is represented by a VIX volatility index.5 Given a particular stochastic discount factor (SDF) I link parameters of risk-neutral dynam-ics of volatility to its historical evolution. Theoretical model implies that risk-neutral volatilitymeasure depends not only on historical structural parameters but also on risk prices. This con-nection logically requires joint estimation using both volatility measures. Garcia et al. (2011)estimate parameters of a continuous-time stochastic volatility model both for objective andrisk-neutral distributions jointly. Risk-neutral measure of volatility is based on option priceseries expansion. In this paper I estimate joint model using the VIX index which is a broaderand likely less noisy measure of volatility. Chernov and Ghysels (2000) use efficient method of 4 See also Renault (2009) 5 See also Jiang and Tian (2007) for detailed justification. 4
  5. 5. moments to estimate jointly historical and risk-neutral distribution parameters and filter outspot volatility. Bollerslev et al. (2011) also estimate joint volatility model but it lacks abovementioned multi-factor volatility specification. Another contribution relative to methodology of Bollerslev and Zhou (2002) is that I keepthe explicit definitions of model innovations in terms of stochastic integrals. This allows me toaccount for all possible interaction between the variables and innovations in analytical form.In particular, Bollerslev and Zhou (2002) rely on unbiased estimator of squared volatility (seeRenault, 2009). Inclusion of the stochastic drift in volatility model somewhat complicates econometric ap-proach. First of all, integrated trend which shows up in discretized model is unobservable andthere is no convenient proxy for it. Hence, I integrate it out which results in higher order ARMAstructure for integrated volatility. But at the same time it preserves identification of structuralmodel parameters and allows for the use of standard GMM procedure (Hansen, 1982). The rest of the paper is organized as follows. Section 2 states the continuous-time stochasticvolatility model of the market return both for historical and risk-neutral distributions. Section 3shows how to discretize continuous-time model and represent it in terms of integrated variables.Section 4 presents the decomposition of volatility premia and shows theoretical contribution ofcentral tendency premia. Section 5 outlines estimation strategy. Section 6 describes empiricalresults. Section 7 concludes.2 The ModelIn this section I present the continuous-time model of the stochastic volatility with the drift thatis also stochastic. This drift represents the persistent central tendency of volatility or its slowlyvarying average level. The continuous-time diffusion is basically the extension of the square-root process used by Heston (1993) for option pricing. I show that with such a modificationmy model has a potential of matching the high persistence of volatility observed in the data asevident from the theoretical autocorrelation function of the spot volatility. I assume a matching square-root form of stochastic discount factor (SDF) that assigns pricesfor shocks both in central tendency and volatility itself. Given the SDF it is easy to bridgehistorical distribution of returns, volatility, and central tendency with its equivalent risk-neutraldistribution used for standard no-arbitrage pricing. Under this equivalent distribution the modelform remains intact but most of the parameters are altered. In particular, assuming negativeprices for volatility and central tendency risk, both processes become more persistent and havea higher unconditional mean level. Consider the probability space (Ω, F, P ) which is a fundamental space of the stochastic 5
  6. 6. market price St . Assume that the log of stock price pt = log St evolves according to thefollowing stochastic differential equation: 1 2 dpt = r + λr − σt dt + σt dWtr , (2.1) 2 2where r is a constant risk-free rate, and λr is equity premium parameter. Here σt plays a role ofinstantaneous variance of the market return. Define the integrated variance of the return overthe h time interval, ˆ ˆ 1 t+h 1 t+h 2 d [p, p]u = σu du ≡ Vt,h . h t h t Also assume that this instantaneous volatility mean reverts to a stochastic central tendencywhich in turn mean reverts to a constant long-term mean of volatility. This volatility structureis similar in spirit to what is suggested by Duffie et al. (2000) in the end of the paper. Theseassumptions may be written in diffusion form as follows: dσt =κσ yt − σt dt + ησ σt dWtσ , 2 2 √ (2.2) dyt =κy (µ − yt ) dt + ηy yt dWty ,where Wtr , Wtσ , and Wty are three standardized Brownian motion processes under the histor- √ical probability measure P . We assume that Wtσ = ρWtr + 1 − ρ2 Wtv or equivalently thatCorr [dWsr , dWsσ ] = ρ which is often called a leverage effect. We also let Wty be completely in-dependent of other shocks. Under the suitable regularity conditions (see Karatzas and Shreve,1997) the above multivariate diffusion has a unique strong solution on R+ . The parameter vec-tor θ is assumed to lie within some compact set Θ ⊂ Rd . Provided that 2µκy ≥ ηy the process 2yt has a stationary Gamma distribution. The reason I can call yt a central tendency is the following. As I show in Section B.2 theautocorrelation of the spot volatility is given by 2 2 2 −1 κσ ηy ηy κσ + κy ησ Corr σt+h , σt = e−κσ h + e−κy h − e−κσ h 2 2 + . κσ − κy κy κy κσ κσThis formula shows that if mean reversion speed of yt is much smaller than the mean reversionspeed of spot volatility itself, then the autocorrelation function in the long horizons will bemainly due to component e−κy h which decays very slowly with small κy . Now let the log stochastic discount factor (SDF) process mt = log Mt be represented by thefollowing SDE: dmt = −rdt − ζr dWtr + ζσ dWtv + ζy dWty . (2.3) 6
  7. 7. Here the vector     ζr λr σt       ζσ   =  − √ 2 (ρλr + λσ ) σt 1   1−ρ √     ζy −λy ytis interpreted as a vector of risk prices arising from different sources of uncertainty. The firstelement of the vector is the price of equity risk, the second is the price of volatility risk, andthe third is the price of risk related to persistent stochastic mean of volatility. Applying Girsanov’s theorem to the model at hand I can define the new set of Brownianmotion processes that are equivalent to the original ˜ dWtr = dWtr + ζr dt, ˜ dWtv = dWtv − ζσ dt, ˜ dWty = dWty − ζy dt.This adjustment in Brownian innovations provides a new set of standard uncorrelated Brownianmotions under risk-neutral probability measure Q on (Ω, F). Under this new probability measure the model may be written as 1 2 ˜ dpt = r − σt dt + σt dWtr , 2 ˜ ˜ 2 ˜ dσt = κσ yt − σt dt + ησ σt dWtσ , 2 (2.4) √ y ˜ µ ˜ ˜ ˜ ˜ d˜t = κy (˜ − yt ) dt + ηy yt dWty , √ ˜ ˜with Wtσ = ρWtr + ˜ 1 − ρ2 Wtv , where the rescaled central tendency is κσ yt = ˜ yt , ˜ κσand the modified parameters are κy κσ κσ κσ = κσ − λσ ησ , ˜ κy = κy − λy ηy , ˜ µ=µ ˜ , ηy = ηy ˜ . ˜ ˜ κy κσ ˜ κσ Note that in general according to Girsanov’s theorem the shift in drift does not alter theinstantaneous diffusion parameters (in this case ησ and ηy ). It may seem that this rule is brokenas the instantaneous diffusion parameter for the central tendency is a multiple of ηy . Thismodification is only due to rescaling of yt itself. This rescaling preserves the interpretation ofmodified yt as a central tendency under the risk-neutral measure. ˜ 7
  8. 8. 3 Exact discretizationClearly, the continuous-time model is a convenient theoretical construct. But in all of theempirical work we only deal with discretely observed data. In this section I show how continuous-time stochastic volatility model with stochastic mean may be exactly discretized. Even afterdiscretization the financial literature does not give a fool-proof recipe to measure spot volatility.It became a de facto standard that we can reliably measure only integrated volatility using highfrequency data. Hence, I proceed in this section by deriving the dynamic discrete model wherestate variables are integrated volatility and central tendency. Besides, for further considerationsof premia it is crucial to look at figures accumulated over some meaningful amount of timerather than instantaneous values. It is well known that the exact discretization of continuous-time square-root process is aheteroscedastic first order autoregression. Since I have two interacting spot variables, the dis-cretized system is likely to be of vector autoregressive form of order one. To make a transferto the integrated state variables I also integrate the innovations which leads to the same orderone vector autoregression but with more complicated moving average innovations of order one. One more complication which will become evident later is the necessity to build the discretemodel for variable integrated over a larger period of time than the lag in the autoregression.In this section I show that the model is not straight vector autoregression anymore but forestimation method of my choice it does not present a significant problem. As Section B.1 shows in more detail the spot volatility model is discretized as σt+h = Aσ σt + Bh yt + Ch + 2 h 2 σ σ σ t,h , (3.1) yt+h = h y Ay yt + Ch + y t,h ,where I define coefficients as κσ Aσ = exp (−κσ h) , h σ Bh = (Ay − Aσ ) , h Ch = µ (1 − Aσ − Bh ) , σ h σ κσ − κy hand Ay = exp (−κy h) , h Ch = µ (1 − Ay ) . y hNote that Ay and Aσ are multiplicative functions of time interval, that is Ay Ay = Ay . Sub- h h u v u+vscripted notation for the error terms means that they are amalgamations of continuous Brow-nian innovations starting from the moment zero to h. In particular, the error structure of the 8
  9. 9. discretized model is given as ˆ t+h ˆ t+h σ √ t,h = ησ σu Aσ σ t+h−u dWu + ηy σ y yu Bt+h−u dWu , t t ˆ t+h y √ t,h = ηy yu Ay y t+h−u dWu . tNote that the volatility innovation accumulates Brownian terms from both central tendencydiffusion, and volatility diffusion itself. Also observe that the system in (3.1) is actually a 2bivariate vector autoregression of order one with state variables (σt , yt ) and heteroscedastic σ yerrors adapted to Ft+h = σ (Wu , Wu | u ≤ t + h). Clearly, the same discretization technique may be applied to the risk-neutral model in (2.4).So I have ˜h 2 ˜ σ ˜ ˜σ σt+h = Aσ σt + Bh yt + Ch + ˜σ , 2 t,h (3.2) yt+h = ˜ ˜ ˜ ˜ Ay yt + C y + ˜y , h h t,hwith ˜h ˜σ ˜ κσ ˜ ˜ ˜σ ˜ ˜ ˜σ Aσ = exp (−˜ σ h) , κ Bh = Ay − Aσ , h h Ch = µ 1 − Aσ − Bh , h κσ − κy ˜ ˜and ˜ Ay = exp (−˜ y h) , κ ˜y ˜ ˜ Ch = µ 1 − Ay . h hThe error structure in the risk-neutral model is ˆ t+h ˆ t+h √ ˜σ t,h = ησ ˜t+h−u dWu + ηy σu Aσ ˜σ ˜ ˜ ˜σ ˜y yu Bt+h−u dWu , t t ˆ t+h √ ˜y = ηy t,h ˜ yu Ay ˜y ˜ ˜t+h−u dWu . t At this stage we already have exactly discretized model of spot volatility and central ten-dency, but my final target is the model for integrated volatility and central tendency. It is alsoargued later that instead of working with instantaneous variables it is more feasible to work 2with integrated variables. In order to make a transfer from instantaneous (σt , yt ) to integratedanalog (Vt,h , Yt,h ) I integrate equations in (3.1) over the reference point in time as a dummy.Integrated volatility and integrated central tendency are defined as ˆ t+h ˆ t+h 1 2 1 Vt,h ≡ σu du, Yt,h ≡ yu du. h t h tHere the first subscripted value denotes the beginning of the time interval, and the seconddenotes the length of this interval. In this particular case the integration interval starts at t 9
  10. 10. and ends at t + h. Integrate the linear system in (3.1) over t as a dummy of integration in the interval [0, h]: ˆ 1 h Vt+h,h = Aσ Vt,h h + σ Bh Yt,h + σ Ch + σ t+s,h ds, h 0 ˆ (3.3) y y 1 h y Yt+h,h = Ah Yt,h + Ch + t+s,h ds. h 0This system of equations is again a bivariate vector autoregression of order one with respectto state vector (Vt,h , Yt,h ). But the error structure is a bit more complicated than it was forinstantaneous state vector. Here the errors aggregate Brownian shocks not only over the period[t + h, t + 2h] but also from the previous period [t, t + h]. This fact makes the error structure amoving average of order one. These errors are measurable with respect to Ft+2h but completelyunpredictable with respect to Ft . Hence the system in (3.3) is of almost VARMA(1,1) form.The error is a composition of shocks over the period [t, t + h] and over [t + h, t + 2h]. But errorterms over both of these periods do not have the same distribution. That is why the systemdoes not comply with the strict definition of VARMA(1,1). Now before writing the integrated version for the risk-neutral model I have to clarify a moregeneral approach to be used. For reasons to be seen in the estimation methodology section itis necessary to integrate the instantaneous discrete system in (3.2) over a larger time intervalthan h. So, denote another positive time variable H ≥ h. the integration gives ˆ ˜h ˜σ ˜ ˜σ 1 H σ Vt+h,H = Aσ Vt,H + Bh Yt,H + Ch + ˜ ε ds, H 0 t+s,h ˆ H ˜ Yt+h,H = ˜y Yt,H + C y + 1 Ah ˜ ˜ εy ds. ˜ h H 0 t+s,h ˜Note the obvious notation Yt,H = κσ Yt,H . This system looks very similar to historical version in κσ ˜(3.3) but with the following important distinctions. First, integration time intervals on the left[t + h, t + h + H] and on the right [t, t + H] clearly overlap. In other words, Vt+h,H and Vt,Hhave some common integrated volatility dynamics in the interval [t + h, t + H].4 Volatility premiaIn this section I derive theoretical implications of the model for the premia related to differentsources of risk. In general, a risk premium is defined as a difference between the objective andrisk-neutral forecasts of the integrated risk factor. There is a large literature dealing with thepremium associated with stochastic volatility. In this paper I hypothesized another source of 10
  11. 11. risk, the shocks in slowly varying average level of volatility. Naturally, I define the premium forthis risk as an excess forecast of the integrated central tendency under objective and risk-neutralprobability measures. The most interesting question now is how this central tendency premiumrelates to volatility premium itself. Thanks to Andersen and Bollerslev (1998) I have a reliable measurement instrument forintegrated volatility. At this point it is beyond the scope of this paper to propose an analogousmeasure for the integrated central tendency. Hence, I will treat this factor as completely unob-servable across the whole paper. Even though it is beyond my reach to quantify the dynamics ofthe volatility central tendency or its premium, I can still say a lot on the basis of the theoreticalmodel. The information I have access to are unconditional moments of volatility and centraltendency premia such as mean, standard deviation, cross-correlations, and autocorrelations. Inthis section I outline the methodology to derive these moments analytically. The method isbased on the representation of time series as infinite integrals with respect to Brownian incre-ments only. This approach together with stochastic calculus makes it very straightforward toderive the moments of interest. Define two premia corresponding to both stochastic volatility and central tendency: V Pt,H = EtQ [Vt,H ] − EtP [Vt,H ] , ˜ CPt,H = EtQ Yt,H − E P [Yt,H ] . tThis means that a volatility risk premium is any excess expected integrated volatility under therisk-neutral measure over the expectation under the historical probability measure. In fact, thepremium is always considered to be the negative of this quantity. It is also widely accepted thatmost of the times the premium associated with stochastic volatility is a negative value. Hence,just for exposition purposes I will consider the negative of this value. The second premium above corresponds to the stochastic central tendency of volatility. Note ˜that the genuine integrated central tendency under the risk-neutral measure is Yt,H = κσ Yt,H . κσ ˜This rescaling is done to justify the use of term “central tendency” in application to the processunder the risk-neutral measure. With this definition it becomes clear how to quantify the importance of the premium as-sociated with shocks in stochastic volatility drift. Define the difference between two premiaas T Pt,H = V Pt,H − CPt,H .This value may be interpreted as a transient premium. The problem now is to characterize these three premia. In the estimation section I willargue that the only two values we observe or at least can find convincing proxies for are realized 11
  12. 12. volatility and VIX volatility index. The first value proxies integrated volatility Vt,h over periodh, and the second proxies risk-neutral expectation of integrated volatility over some largerperiod H. In this paper I eliminate any possibility of an error by leaving these values intactand not doing any forecasting of historical volatility. Forecasting approach was taken by Eraker(2009) with a very simple lagged realized volatility, Bollerslev et al. (2010) with HAR-RV, orTodorov (2010) with VAR-based forecast. This limitation immediately moves all premia intothe rank of unobservables. On the other hand, my model allows for analytical expressions forvarious analytical moments of premia. Below I briefly outline the approach I take to derivethese moments. The proof with all necessary details is given in Section B.5. First of all I represent both spot volatility and central tendency as an infinite stochasticintegrals with respect to Brownian motions only. Under the historical measure the modelbecomes ˆ t √ yt = µ + ηy yv Ay dWv , t−v y −∞ ˆ t ˆ t √ σ y 2 σt = µ + ηy yv Bt−v dWv + ησ σv Aσ dWv . t−v σ −∞ −∞Note that spot volatility accumulates shocks of both Brownian motions. Integrating this ex-pression over the period [t, t + H] I obtain another representation of integrated volatility andcentral tendency as infinite integrals: ˆ t+H ˆ t+H ˆ t ˆ t+H √ √ HYt,H =µH + ηy yv Ay du u−v y dWv + ηy yv Ay du dWv , u−v y t v −∞ t ˆ t ˆ t+H ˆ t+H ˆ t+H √ σ y √ σ y HVt,H =µH + ηy yv Bu−v du dWv + ηy yv Bu−v du dWv −∞ t t v ˆ t ˆ t+H ˆ t+H ˆ t+H + ησ σv Aσ du dWv + ησ u−v σ σv Aσ du dWv . u−v σ −∞ t t vNote that each expression above naturally breaks down into two parts. One accumulates randomall shocks up to time t and the other shocks from t to t + H only. Hence, taking expectationswith respect to historical measure and information up to time t leaves only the first part ofstochastic integrals intact. The second part is completely unpredictable with respect to Ft . The representation under the risk-neutral measure is very similar except to slight changeof notation. Similarly, after taking expectation with respect to the measure Q I obtain the ˜ ˜infinite integrals up to time t with respect to Brownian motions W σ and W y . In order to havea meaningful expression for the risk premia I have to measure these Brownian motions underthat same measure I used for W σ and W y . This means that I have replace Brownian shocksunder the risk-neutral measure by their equivalents under the physical measure. At this pointit becomes meaningful to take the difference between risk-neutral and historical expectations of 12
  13. 13. volatility and central tendency. In the end I can show that the three above defined premia may be represented as ˆ t P √ C y CPt,H = E [CPt,H ] + yv v,t dWv , −∞ ˆ t ˆ t P σ √ V y V Pt,H = E [V Pt,H ] + σv ωv,t dWv + yv v,t dWv , −∞ −∞ ˆ t ˆ t √ T Pt,H = E P [T Pt,H ] + σ σv ωv,t dWv + yv T y v,t dWv . −∞ −∞The first term in each expression above is an unconditional mean of the corresponding premium.The rest are stochastic integrals with respect to Brownian motion increments. Functions v,tand ωv,t are completely deterministic and only depend on structural parameters of the model. Unconditional means are ˆ ˆ t+H P µ t κσ ˜ σ ˜ E [V Pt,H ] = (˜ − µ) − µ λy ηy Bu−v + λσ ησ Aσ u−v dudv, H −∞ t ˜ κσ ˆ ˆ t+H µ t κσ ˜ P E [CPt,H ] = (˜ − µ) − µ λy ηy Ay dsdv, H −∞ t κσ s−v ˜ ˆ t ˆ t+H µ κσ ˜ σ ˜s−v + λσ ησ Aσ ˜s−v dsdv. E P [T Pt,H ] = − λy ηy Bs−v − Ay H −∞ t ˜ κσNote that both volatility and central tendency premia have one component in common, (˜ − µ) H. µIf I normalize all premia by the length of the time interval, the common component is simplythe difference between long-term mean of volatility, both spot and integrated, under two equiv-alent measures. This means that on average the volatility premium is not simply the differencebetween unconditional means of realized volatility and risk-neutral volatility measure. Finally, using the representation above it is quite simple to compute unconditional momentsof the three premia. For example, the variance of the premia are the following deterministicintegrals: ˆ t P C 2 V [CPt,H ] = µ v,t dv, −∞ ˆ t V 2 V P [V Pt,H ] = µ (ωv,t )2 + v,t dv, −∞ ˆ t T 2 V P [T Pt,H ] = µ (ωv,t )2 + v,t dv. −∞Correlations between premia and autocorrelations are derived analogously. Since the above moments are hard to analyze analytically I will proceed to analyze them em- 13
  14. 14. pirically by substituting estimates of structural parameters and using their variance-covariancematrix to compute standard errors by delta method.5 EstimationIn this section I describe how to jointly estimate parameters of continuous-time stochasticvolatility model under the historic distribution in (2.2) and under the risk-neutral distributionin (2.4) with limited information. The first subsection deals with a substantial hurdle for a financial econometrics, namely themeasurement of unobservable factors such as volatility and introduced here central tendency.The first problem is how to measure volatility itself under two different probability measures.Under the historical measure Andersen and Bollerslev (1998) suggested6 to use intra-day highfrequency data on the returns. They show that with an interval going to zero the almost surelimit of sum of squared returns is an integrated volatility. The daily realized volatility measureis readily reported and accessible from multiple sources. Under the risk-neutral measure thevolatility is one of the main factors determining option prices. Using this fact Britten-Jonesand Neuberger (2000) theoretically justify the use of a large set of option prices to construct arisk-neutral measure of volatility manifested in VIX volatility index. A slight technical problem with these two data series is that they do not match with respectto the integration horizon of volatility. Realized volatility is a proxy for daily integrated volatil-ity, and VIX is based on options with maturity of one month (22 business days). Moreover,realized volatility is a genuine volatility measure while VIX is a risk-neutral expectation of fu-ture integrated volatility. It is easy to synchronize these data by aggregating realized volatilityover a month period but it is far more speculative to form its forecast. Different approacheslead to slightly different results (see Eraker, 2009; Todorov, 2010). For this paper I will stay outof this debate and will only use what I reliably have in the data and adjust econometric modelappropriately. This adjustment amount to augmenting the model innovation with the forecasterror of conditional expectation of volatility and central tendency. This approach reformulatesthe model in terms of conditional forecasts as state variables. In Section 3 I have already derived the vector autoregression model for both historical andrisk-neutral distributions where two factors are integrated volatility and central tendency. As Ialready stated I do not propose to measure the central tendency but rather treat it as unobserv-able and work around this complication. The trick to get rid of the integrated central tendencyin the model is to marginalize it. By doing so with VAR(1,1) type of model I marginalize it intoARMA(2,2) model. This trick works for the realized volatility which is measured over one day 6 See also Meddahi (2002) 14
  15. 15. and is lagged one day in the model. For the risk-neutral model the exact order of the movingaverage component is a bit hard to pin but, again, it does not play a significant role as long asI know the analytical structure of the innovations. In the second part of this section I present the conditional moment I use to set up GeneralizedMethod of Moments estimator (Hansen, 1982). It turns out that the GMM estimation ofARMA-type models is very unstable if used to estimate all structural parameters including theunconditional mean. To circumvent this problem I use the two step approach. First, I estimatethe unconditional means of both historical and risk-neutral measures of volatility. This approachis known as variance targeting in GARCH literature and justified by Horvath et al. (2006) andFrancq et al. (2009). On the second step I treat the mean as given and estimate the rest ofthe parameters by GMM. For that I compute analytically conditional mean and conditionalsecond moment of integrated volatility explicitly taking into account all possible correlationsbetween model innovations. The first moment only identifies speed of mean reversion parameter.The second moment is necessary to identify instantaneous diffusion parameters. In total thisapproach gives me four moments, two for each probability measure. As instruments I use laggedrealized volatility, VIX, daily market return, and squared return.5.1 MeasurementClearly, the first problem an econometrician faces with these types of models is that theyare formulated for the variables that are observed at best over discrete time intervals. Toovercome this first obstacle I have derived the exact discretization of the same model. Since Iam interested in estimating parameters under both historical and risk-neutral distributions, Ihave two discretized models, (3.1) and (3.2). 2 Now the next problem is that volatility σt and its stochastic mean yt are not observablevariables even in discrete time intervals. Since point-in-time volatility is unobserved Andersenand Bollerslev (1998) suggested to estimate integrated volatility using high frequency observedreturn data. In particular, the following convergence result provides the basis for such estima-tion: ˆ n t+h 2 a.s. 2 RVt,h ≡ rt+ j−1 h,t+ j h −→ σu du ≡ Vt,h . n n j=1 tSince we can more or less reliably observe only integrated volatility, but not its point-in-timevalue, I have to resort to the methodology of Bollerslev and Zhou (2002). They transform allknown results into relations between integrated volatility and apply GMM for estimation. Onceagain, I will adopt the measurement of integrated volatility under the physical measure P bythe realized volatility. Given discrete time observations on the theoretically reliable proxy of integrated volatility 15
  16. 16. it becomes natural to transform the model into (3.3) which is of VARMA(1,1) form with het-eroscedastic errors and state vector [Vt,h , Yt,h ]. Another difficulty is that in the VARMA(1,1)model (3.3) the integrated central tendency Yt,h is not observable. Even worse, there is nogood proxy for this variable that I am aware of. Besides, the purpose of this paper is not tosuggest a measure for this unobservable component but to circumvent this problem all togetherand estimate the model with what we reliably have in the data. To circumvent the absence ofa good proxy for a latent variable I marginalize the observable and compute the moments interms of what is known in the data. For more details see the following Section 5.2. At this point I have formulated the model under the objective measure P for the integratedvolatility which may be accurately proxied by the realized volatility. Realized volatility isnormally constructed from the intra-day data to obtain a measure on a daily frequency, that ish = 1. In order to estimate the risk-neutral model parameters I need the data on integrated volatilityunder the measure Q. Britten-Jones and Neuberger (2000) provide the result that connectsoption data with risk-neutral volatility forecast. Specifically, ˆ ∞ C (t + H, K) − max (St − K, 0) V 2 IXt,H =2 dK = EtQ [Vt,H ] , 0 K2where C (t + H, K) is the price of call option maturing at time t + H with strike price K, andmax (St − K, 0) is the intrinsic value of this option at time t. If H is set at one month or 22days period, then this expression is well proxied by the VIX index of volatility. Note that I deliberately used time length H rather than h to stress the point that VIX is aforecast of integrated volatility over a period of 22 days rather than one days in case of realizedvolatility. So, even though VIX as a proxy for EtQ [Vt,H ] is observed on a daily basis, it gives aprediction of volatility over 22 days in the future. In order to account for that in the theoreticalmodel I integrate discrete point-in-time risk-neutral model in (3.2) over a period of time Hrather than h as in the case of historical model. Even though this discrete-time model is not ofsimple familiar structure it is still manageable in terms of computing analytical moments andapplying GMM procedure.5.2 Moment conditionsIn this section I compute analytically the first two moment conditions both for objective andrisk-neutral measures of integrated volatility. The first step is to eliminate unobservables from econometric model or, in other words,marginalize the observed variable represented by integrated volatility. Using the lag operator L 16
  17. 17. and taking the conditional expectation this system may be represented as EtP [(1 − Aσ L) Vt+h,h ] = Bh EtP [Yt,h ] + Ch , h σ σ EtP [(1 − Ay L) Yt+h,h ] = h y Ch .Multiply the first equation by (1 − Ay L), shift the time by h, and make a substitution from the hsecond equation to obtain EtP [(1 − Ay L) (1 − Aσ L) Vt+2h,h − M1 ] = 0, h hwith M1 ≡ (1 − Ay ) (1 − Aσ ) µ. h h Analogously to the historical model I can marginalize integrated volatility and formulatethe first moment completely analogously with a simple change in notation. One more problemI see here is that integrated volatility Vt,H under the risk-neutral distribution is not directlyobservable. What we have in the data is only a proxy for its risk-neutral forecast EtQ [Vt,H ] Qgiven by VIX volatility index. So, introduce a new observable variable Vt,H : Vt,H ≡ EtQ [Vt,H ] . QApplying the law of iterated expectations this replacement of the latent variable does notactually change the moment: EtQ ˜ 1 − Ay L ˜ Q ˜ 1 − Aσ L Vt+2h,H − M1 = 0, h hwith ˜ ˜ M1 ≡ 1 − Ay ˜ ˜ 1 − Aσ µ. h h The first moments identify only mean and speed parameters of diffusion. In order to identifyinstantaneous variance and price of shocks one has to compute additional moments. As I showin Section B.3, the second conditional moments of integrated volatility and spot variables may 17
  18. 18. be written as 2 EtP (hVt,h )2 = A1 σt + A2 yt + A3 + aσ σt + bσ yt + cσ 2 h 2 h h , 2 EtP σt+h 4 = a1 σt + a2 yt + a3 + Aσ σt + Bh yt + Ch 2 h 2 σ σ , EtP σt+h yt+h 2 = c2 yt + c3 + Aσ σt + Bh yt + Ch (Ay yt + Ch ) , h 2 σ σ h y EtP yt+h 2 = b2 yt + b3 + (Ay yt + Ch )2 , h y EtP σt+h 2 = Aσ σt + Bh yt + Ch , h 2 σ σ EtP [yt+h ] = Ay yt + Ch . h y 2The only observable here is integrated volatility Vt,h . The rest are latent variables which canbe eliminated by taking appropriate lags and making substitutions. For example, the spotvolatility and central tendency equations may be written as EtP [(1 − Ay L) yt+h ] = Ch , h y Et (1 − Aσ L) σt+h h 2 σ σ = Bh yt + Ch .Multiply the last equation by (1 − Ay L), shift the time by h using the law of iterated expecta- htions, and finally substitute the first equation in to get EtP (1 − Ay L) (1 − Aσ L) σt+2h = M1 . h h 2 4 2The expression for the second moment of integrated volatility includes spot variables σt , σt yt , 2 2yt , σt , and yt . Using the above approach each one of these variables is eliminated with the endresult of EtP 1 − (Aσ )2 L (1 − Aσ Ay L) 1 − (Ay )2 L (1 − Ay L) (1 − Aσ L) Vt+5h,h = M2 , h h h h h h 2where constant M2 is defined in Section B.3 and its definition involves instantaneous varianceparameters. Parameters in the first and second moments of integrated volatility under two probabilitymeasures are implicitly given by the vectors (µ, κσ , κy , ησ , ηy ) and (˜, κσ , κy , ησ , ηy ), respectively. µ ˜ ˜ ˜The connection between these parameters is κy κσ κσ κσ = κσ − λσ ησ , ˜ κy = κy − λy ηy , ˜ µ=µ ˜ , ηy = ηy ˜ . ˜ ˜ κy κσ ˜ κσ 18
  19. 19. If I estimate parameters of the two models jointly, then the parameter vector becomes7 θ = (µ, κσ , κy , ησ , ηy , λσ , λy ) ,which differs from historical set of parameters only by risk prices of volatility and central ten-dency innovations, λσ and λy . Both first and second moments are expressed given fine information set Ft which contains 2all past observations of point-in-time volatility σt and yt . Clearly, this information is notavailable to econometrician. Hence, an additional technical step of reduction in information setis necessary (Meddahi and Renault, 2004). Coarser information set includes only observationson past integrated volatilities. Finally, the applicability of GMM is argued in Section B.6. There I claim that the momentrestrictions, model innovations, integrated variables, their interactions are reducible to a simplestochastic integral. Consequently, for this variable I show the existence of finite fourth momentand central limit theorem.6 ResultsIn this section I will describe the data I use for my empirical exercise. Then I will outline theresults of estimation of continuous-time model parameters. The rest of the section is dedicatedto analyzing the main result of the paper. This result states that the volatility premium ismostly composed of the premium on central tendency. Part of it is composed of the premiumon shocks that are relatively more volatile but short-lived. In other words, the additional sourceof risk, the one that drives volatility around its persistent stochastic drift, does bear somesmall but statistically significant premium. It is also interesting to note that the instantaneousvariance of a central tendency is only a small fraction of instantaneous variance of additionalvolatility shocks. Hence, small shocks that preserve its effect long into the future are priced andmake the largest portion of compensation. In this study I use the following data. Daily volatility index, VIX, is constructed by CBOE8for the period starting from 1990. This index proxies the integrated volatility forecast over thefuture 22 business days. Daily S&P500 index prices, SPX, and realized volatility measure, RV,are reported by Oxford-Man Institute9 for the period starting from 1996. The data for marketindex and log daily return are shown in Figure 1 on page 28. Both volatility indexes are shown 7 Due to postulated causality links in the model, estimation of leverage parameter and equity premium doesnot affect estimation of volatility parameters. But for completeness, I estimate the full joint model. For derivationof return moments see Section B.4. 8 CBOE, VIX Historical Price Data, 9 Oxford-Man Institute’s “realized library”, 19
  20. 20. in Figure 2 on page 28. I report descriptive statistics of the data in Table 1 on page 29. Inaddition, estimates of autocorrelation function are shown in Figure 3 on page 29. The results of several models estimation are given in Table 2 on page 30. The first fourcolumns in this table correspond to a benchmark estimation of the univariate square-root dif-fusion to the daily realized volatility and VIX separately. The next four columns correspond toestimation of bivariate diffusion model with central tendency to RV and VIX separately. Thelast column is main estimation result of the paper as it fits the model with central tendencyjointly to RV and VIX volatility measures. I used heteroscedasticity and autocorrelation robustestimator of moments covariance matrix (Newey and West, 1987) with Bartlett kernel and 5lags for univariate models, and 50 lags for bivariate models. The instruments for the estimationin benchmark models are lagged variables themselves. For the main model estimation I usedlags of realized volatility, VIX volatility index, and squared daily log-return. First of all from Table 2 on page 30 it is clear that none of the models, except for the one-component model for realized volatility, is not rejected by the J-test with p-value varying from57% to 76% for benchmark cases, and 91% for the main model. In all benchmark cases theparameters are highly significant with an exception of diffusion parameter of central tendency. The very first column suggests that the mean reversion of realized volatility is 0.6227. Thisnumber corresponds to half-life10 of a shock to a little under half a day. This does not lieanywhere close to roughly 15 days evident from estimated autocorrelation function in Figure 3on page 29. The estimated mean reversion speed of VIX is 0.4996 which corresponds to half-lifeof a little over half a day which also does not match about 45 days evident from the figure. Thisobservation clearly makes univariate models inadequate in view of the real data. One can see that the modeling of the central tendency is relevant since both κy and ηy aresignificantly different from zero. Bivariate models bring an additional degree of freedom byintroducing and additional shock which should cover the fat tails feature of the data. This isthe most clearly seen in case of realized volatility. There the speed of mean reversion of thecentral tendency is estimated as 0.0255 which corresponds to almost 12 days of half-life. Eventhough it still does not exactly reach the level seen in the data, it is a definite improvementin the fit of the model. The central tendency mean reversion estimate for the VIX series is0.0123 which corresponds to 25 days of half-life. This number is substantially larger than in theunivariate model but still falls short of empirical 45 days. Besides mean reversion, one can note some interesting patterns in diffusion parameter esti-mates. In univariate case the diffusion parameter of realized volatility is estimated as 0.0707, while in bivariate case two diffusion estimates are 0.0833 and 0.0125. This means that theshocks changing the persistent drift are much less volatile than shocks changing the movement 10 log 2 Half-life of a shock is defined as T = κ , where κ is the long-term speed of mean reversion. 20
  21. 21. of volatility around its long-term drift. The univariate estimate is somewhere in the middle.The same picture is observed in the case of VIX. Besides, diffusion parameter estimates aresomewhat smaller for VIX than they are for RV. The estimation of the joint model for RV and VIX gives us estimates of all parameters in themodel, including leverage parameter and equity risk price. Central tendency estimates are stillvery significant. The main benefit of joint model estimation though are the risk prices. Pointestimates of λσ and λy are 0.2013 and 1.0929, respectively. This means that central tendencyshocks are five times more expensive. The model estimates imply that the risk-neutral speed of mean reversion of central tendencyis 0.0178 − .2013 × 1.0929 = .0098, and .8989 − .1041 × .2013 = .8780 which are not too farfrom estimates of VIX model. The long-term speed of mean reversion of 0.0098 correspondsto 70 days of half-life which is actually more than 45 days in the data. Hence, the shock ontop of the central tendency is very very short lived but several time more volatile. Both ofthese factors contribute to a higher unconditional mean of volatility in the risk-neutral world,21% against 13% daily in annualized standard deviation units. This observation is due tothe identity µκy κσ = µκy κσ which says that all else equal the higher unconditional mean of ˜˜ ˜risk-neutral volatility requires higher persistence of either central tendency or volatility or both. Given the estimates of model parameters I plug them in to analytical expressions for un-conditional moments of volatility premium, central tendency premium, and their difference. Byplugging in the estimates of these parameters I obtain point estimates of unconditional momentsof unobservable risk premia. Standard errors for these estimates are computed using the delta ˆmethod. Having the estimates θ of structural parameters and their covariance matrix Ω it is ˆeasy to compute estimates of γ = f (θ) and their standard errors. Point estimates are γ = f θ ˆ ˆand the covariance matrix is computed through delta method and given by ∂f ˆ ˆ ∂f θ ˆ V (ˆ ) = γ θ Ω . ∂θ ∂θ Implications for unconditional moments of volatility are given in Figure 4 on page 31 throughFigure 7 on page 33. Risk premia mean and standard deviation are given in Figure 9 on page34 and Figure ?? on page ??. These graphs, except for autocorrelations, plot unconditionalmoments over forecasting horizon H which I vary from 1 to 22 business days. The last number isthe same as used in computing VIX volatility index. Everything below that is a shorter intervalfor volatility integration and is used to reveal the relative importance of central tendency overdifferent forecasting horizons. The implied estimates of autocorrelation function both for RV and VIX compared with thedata are displayed in Figure 4 on page 31. These plots show that both implied autocorrelations 21
  22. 22. somewhat underestimate empirical counterparts. At the same time the long-term decay iscaptured very close and even overstated for the risk-neutral series. Hence, the main reasonfor the disparity are the overestimated diffusion parameters which contribute to large impliedvariation in stochastic volatility and lower autocorrelation function across the board. Figure 5 on page 32 shows that standard deviation of volatility, central tendency, and theirdifference. There is no need for unconditional mean picture since it is known to be a constantgiven in Table 2 on page 30. The plot shows standard deviation normalized by a period length.There we see that the volatility and central tendency standard deviations are between 4%and 6% daily. It is interesting to note that with increase in aggregation interval the standarddeviation decreases only marginally. This is due to high persistence of both series. The standarddeviation of the difference is above 7% for 1 day of aggregation and below 3% for the upper end.This says that the transitory shocks in volatility are much more evident on the short forecastinghorizons although still by order smaller than the effects of persistent component. Figure 7 on page 33 shows the estimates of correlations between volatility and its componentsover different aggregation intervals. The most evident result is that the volatility itself is highlycorrelated with its central tendency. On the longer horizons the point estimate is very close toone. The correlation between volatility and its transitory innovations is much more pronouncedat shorter aggregation intervals, above 50%, where transitory shocks are more noticeable asjudged by its daily standard deviation. As expected due to assumption of uncorrelated Brownianinnovations, the correlation between central tendency and transitory shocks is virtually zero. Figure 7 on page 33 plots autocorrelations of three series for lags up to 90 days and one dayof aggregation. The most interesting observation is that the integrated volatility autocorrelationsomewhat understates the empirical estimates of autocorrelation of realized volatility given inFigure 3 on page 29. This observation is a positive reality check of the estimation results. Nat-urally, the autocorrelation of central tendency is even higher. And there is not much persistencein the difference between the two series. Figure 8 on page 33 is exactly analogous to the previous picture with the only difference thatthe integration interval is now 22 business days. Here we see that the volatility autocorrelationis much closer to the central tendency due to longer aggregation of the series. Figure 9 on page 34 shows the most striking result. On this plot I present the comparisonbetween unconditional means of volatility premium, central tendency premium, and their differ-ence in annualized variance units. The point estimates unambiguously suggest that the impliedcentral tendency premium is actually larger on average than the volatility premium itself overall considered forecasting horizons. Volatility premium point estimate goes from under 1% perday to roughly 2.5% per day for 22 days of integration. Central tendency point estimates areeverywhere above. Confidence intervals (95%) for these two premia are approximately plus 22
  23. 23. minus 0.5% daily. The actual test for significant difference between the two premia is given by analyzingconfidence intervals of the mean transitory premium. Point estimate is around 0.3% everywhere.The 95% confidence interval does not include zero. This suggests that the difference betweenthe two premia is significant for forecasting horizons of several days and above. Hence, I can saythat the unconditional means of the two premia are distinguishable for more than few days in thefuture. This also implies that the transitory shocks contribute a statistically significant thoughsmall weight into risk compensation. The basis for this result is evident from the previouslyconsidered graphs and estimation results. There it was clear that volatility and central tendencyare very similar to each other. But the shocks that come on top of the central tendency arevery volatile even though very short lived. This seems to be enough to make a contribution tothe volatility premium unconditional level. To conclude this section I want to stress the main result which says that the volatilitypremium on intervals longer than a few days is mainly due to compensation of highly persistentshocks that drive stochastic drift of volatility.7 ConclusionIn this paper I proposed the continuous-time stochastic volatility model with varying centraltendency. As a main result of the paper I argue that the major part of volatility risk premium isdue to compensation for highly persistent shocks in volatility, those that drive central tendency.Additional short lived but very volatile shocks that drive volatility around its central tendencyare associated with a small but significant negative premium. My approach has several very conservative limitations. First, I treat central tendency ascompletely unobservable and for the purposes of estimation integrate it out. It would be a verypromising avenue of future research to devise a reliable measure of integrated central tendencyanalogous to integrated volatility. The second limitation is that I do not take a risk of speculationand do not propose a specific methodology to form a conditional forecast of historical volatility.This forecast could allow me to identify explicitly volatility premium. On the other hand, if Ileave central tendency as unobservable the value of computing volatility premium by itself isdoubtful. Hence, in order to gain insight into joint dynamics of volatility and central tendencypremia it would take a substantial theoretical work. In spite of these limitations my approachdoes not preclude analytical computation of unconditional moments of different premia andtheir difference. 23
  24. 24. ReferencesAdrian, T. and Rosenberg, J. (2008). Stock Returns and Volatility: Pricing the Short-Run and Long-Run Components of Market Risk. Journal of Finance, 63(6):2997–3030.Andersen, T. G. and Bollerslev, T. (1997). Heterogeneous Information Arrivals and Return Volatility Dynamics: Uncovering the Long-Run in High Frequency Returns. Journal of Fi- nance, 52(3):975.Andersen, T. G. and Bollerslev, T. (1998). Answering the Skeptics: Yes, Standard Volatility Models do Provide Accurate Forecasts. International Economic Review, 39(4):885–905.Andersen, T. G. and Lund, J. (1997). Stochastic Volatility and Mean Drift in the Short Rate Diffusion: Sources of Steepness, Level and Curvature in the Yield Curve.Andrews, D. W. K. (1988). Laws of Large Numbers for Dependent Non-Identically Distributed Random Variables. Econometric Theory, 4(3):458–467.Balduzzi, P., Das, S. R., and Foresi, S. (1998). The Central Tendency: A Second Factor in Bond Yields. Review of Economics and Statistics, 80(1):62–72.Bansal, R. and Yaron, A. (2004). Risks for the long run: A potential resolution of asset pricing puzzles. Journal of Finance, 59(4):1481–1509.Bates, D. S. (2000). Post-’87 crash fears in the S&P 500 futures option market. Journal of Econometrics, 94(1-2):181–238.Bollerslev, T., Gibson, M., and Zhou, H. (2011). Dynamic estimation of volatility risk premia and investor risk aversion from option-implied and realized volatilities. Journal of Economet- rics, 160(1):235–245.Bollerslev, T. and Mikkelsen, H. O. (1996). Modeling and pricing long memory in stock market volatility. Journal of Econometrics, 73(1):151–184.Bollerslev, T., Tauchen, G., and Sizova, N. (2010). Volatility in Equilibrium: Asymmetries and Dynamic Dependencies. SSRN Electronic Journal.Bollerslev, T., Tauchen, G., and Zhou, H. (2009). Expected Stock Returns and Variance Risk Premia. Review of Financial Studies, 22(11):4463–4492.Bollerslev, T. and Zhou, H. (2002). Estimating stochastic volatility diffusion using conditional moments of integrated volatility. Journal of Econometrics, 109:33 – 65. 24
  25. 25. Britten-Jones, M. and Neuberger, A. (2000). Option Prices, Implied Price Processes, and Stochastic Volatility. Journal of Finance, 55(2):839–866.Carr, P. and Wu, L. (2009). Variance Risk Premiums. Review of Financial Studies, 22(3):1311– 1341.Chernov, M., Gallant, R. A., Ghysels, E., and Tauchen, G. (2003). Alternative models for stock price dynamics. Journal of Econometrics, 116(1-2):225–257.Chernov, M. and Ghysels, E. (2000). A study towards a unified approach to the joint estima- tion of objective and risk neutral measures for the purpose of options valuation. Journal of Financial Economics, 56:407–458.Christoffersen, P., Jacobs, K., Ornthanalai, C., and Wang, Y. (2008). Option valuation with long-run and short-run. Journal of Financial Economics, 90(3):272–297.Corsi, F. (2009). A Simple Approximate Long-Memory Model of Realized Volatility. Journal of Financial Econometrics, 7(2):174–196.Cox, J. C., Ingersoll, J. E., and Ross, S. A. (1985). A Theory of the Term Structure of Interest Rates. Econometrica, 53(2):385–407.Drost, F. C. and Nijman, T. E. (1993). Temporal Aggregation of Garch Processes. Econometrica, 61(4):909–927.Duffie, D., Pan, J., and Singleton, K. J. (2000). Transform Analysis and Asset Pricing for Affine Jump-Diffusions. Econometrica, 68(6):1343–1376.Engle, R. F. and Lee, G. G. J. (1996). Estimating Diffusion Models of Stochastic Volatility. In Rossi, P. E., editor, Modeling Stock Market Volatility: Bridging The Gap to Continuous Time, pages 333–355. Academic Press.Engle, R. F. and Lee, G. G. J. (1999). A Long-Run and Short-Run Component Model of Stock Return Volatility. In Engle, R. F. and White, H., editors, Cointegration, Causality, and Forecasting. A Festschrift in Honour of Clive W. J. Granger, pages 475–497. Oxford University Press.Epstein, L. G. and Zin, S. E. (1989). Substitution, Risk Aversion, and The Temporal Behavior of Consumption and Asset Returns: A Theoretical Framework. Econometrica, 57(4):937–969.Eraker, B. (2009). The Volatility Premium. 25
  26. 26. Francq, C., Horvath, L., and Zakoian, J.-M. (2009). Merits and drawbacks of variance targeting in GARCH models.Gallant, R. A., Hsu, C.-T., and Tauchen, G. (1999). Using Daily Range Data to Calibrate Volatility Diffusions and Extract the Forward Integrated Variance. Review of Economics and Statistics, 81(4):617–631.Garcia, R., Lewis, M.-A., Pastorello, S., and Renault, E. (2011). Estimation of objective and risk-neutral distributions based on moments of integrated volatility. Journal of Econometrics, 160(1):22–32.Hansen, L. P. (1982). Large Sample Properties of Generalized Method of Moments Estimators. Econometrica, 50(4):1029–1054.Heston, S. L. (1993). A Closed-Form Solution for Options With Stochastic Volatility With Applications to Bond and Currency Options. Review of Financial Studies, 6:327–343.Horvath, L., Kokoszka, P., and Zitikis, R. (2006). Sample and Implied Volatility in GARCH Models. Journal of Financial Econometrics, 4(4):617–635.Jiang, G. J. and Oomen, R. C. A. (2007). Estimating Latent Variables and Jump Diffusion Models Using High-Frequency Data. Journal of Financial Econometrics, 5(1):1–30.Jiang, G. J. and Tian, Y. S. (2007). Volatility from Option Prices : An Examination of the VIX Index. Journal of Derivatives, 14(3):35–60.Karatzas, I. and Shreve, S. E. (1997). Brownian Motion and Stochastic Calculus. Springer, Berlin.Meddahi, N. (2002). A theoretical comparison between integrated and realized volatility. Jour- nal of Applied Econometrics, 17(5):479–508.Meddahi, N. and Renault, E. (2004). Temporal aggregation of volatility models. Journal of Econometrics, 119(2):355–379.Newey, W. K. and West, K. D. (1987). A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix. Econometrica, 55(3):703–708.Renault, E. (2009). Moment-Based Estimation of Stochastic Volatility Models. In Andersen, T. G., Davis, R. A., and Kreib, J.-P., editors, Handbook of Financial Time Series, volume 57, pages 269–311. Springer. 26
  27. 27. Reschreiter, A. (2010). Inflation and the Mean-Reverting Level of the Short Rate. The Manch- ester School, 78(1):76–91.Reschreiter, A. (2011). The effects of the monetary policy regime shift to inflation targeting on the real interest rate in the United Kingdom. Economic Modelling, 28(1-2):754–759.Tauchen, G. (2005). Stochastic Volatility in General Equilibrium.Todorov, V. (2010). Variance Risk-Premium Dynamics: The Role of Jumps. Review of Financial Studies, 23(1):345–383.White, H. (1984). Asymptotic Theory for Econometricians. Academic Press, Orlando, Florida. 27
  28. 28. A Tables and Figures S&P500 index 1600 1400 SPX 1200 1000 800 600 400 S&P500 log returns 15 10 logR 5 0 −5 −10 97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 19 19 19 20 20 20 20 20 20 20 20 20 20 20 20 date Figure 1: Daily S&P500 index (SPX) and market log return (logR). Volatility measures 140 120 RV 100 80 VIX 60 40 20 0 Difference 60 40 RV-VIX 20 0 −20 −40 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 199 199 199 200 200 200 200 200 200 200 200 200 200 201 201 date Figure 2: Daily option-based volatility index (VIX), and realized volatility (RV). 28
  29. 29. Min Max Mean Std Skewness Exc. Kurtosis 100*log(R) -9.47 10.96 0.019 1.31 -0.20 7.42 VIX 9.89 80.86 22.03 8.62 1.98 6.87 RV 2.38 118.75 13.34 8.36 3.34 19.87Table 1: Descriptive statistics for market log returns (logR), option-based volatility index (VIX),realized volatility (RV), and their difference, volatility risk premium (VRP). Autocorrelation function 1.0 VIX 0.8 RV logR 0.6 0.4 0.2 0.0 −0.2 0 10 20 30 40 50 60 70 80 90 Lags, daysFigure 3: Autocorrelation function for option-based volatility index (VIX), realized volatility(RV), and log market return (logR). 29
  30. 30. θ RV VIX RV VIX RV,VIX µ 0.0032 (0.0003) 0.0068 (0.0004) 0.0040 (0.0006) 0.0069 (0.0017) 0.0046 (0.0005) κσ 0.6227 (0.0320) 0.4996 (0.0226) 0.8961 (0.0110) 0.8998 (0.0039) 0.8989 (0.0057) ησ 0.0707 (0.0106) 0.1391 (0.0048) 0.0833 (0.0010) 0.1485 (0.0008) 0.1041 (0.0225) κy 0.0255 (0.0089) 0.0123 (0.0036) 0.0178 (0.0038) ηy 0.0125 (0.0070) 0.0054 (0.0042) 0.0073 (0.0033) λσ 0.2013 (0.0786) λy 1.0929 (0.4835) λr 0.5156 (0.0146) ρ 0.0456 (0.0214) J 10.20 2.60 2.01 1.56 8.34 p 0.07 0.76 0.57 0.67 0.91 df 5 5 3 3 15Table 2: Estimation results of the model. The parameters have the following interpretation:µ is the unconditional mean of historical volatility; κσ and κy are mean reversion speed pa-rameters for volatility and central tendency, respectively, under historical measure; ησ and ηyare instantaneous diffusion parameters; λσ and λy are risk prices. Risk-neutral speeds of meanreversion are κσ = κσ − λσ ησ and κy = κy − λy ηy . Risk-neutral volatility mean is given by ˜ ˜µ = κσ κy µ. Instantaneous correlation between return and volatility shocks is ρ, and the equity˜ κσ κy ˜ ˜risk price is λr . 30
  31. 31. RV True 0.8 Implied +/− 2σ Autocorrelation 0.6 0.4 0.2 0 −0.2 10 20 30 40 50 60 70 80 90 Lag, days VIX 0.8 Autocorrelation 0.6 0.4 0.2 0 −0.2 10 20 30 40 50 60 70 80 90 Lag, daysFigure 4: Autocorrelation function for option-based volatility index (VIX), realized volatility(RV), and log market return (logR). 31
  32. 32. 12 V[V] Standard deviation of the volatility, % V[Y] 10 V[V−Y] 8 6 4 2 0 5 10 15 20 Forecast horizon, daysFigure 5: Implied standard deviations of daily volatility, V[V], central tendency, V[Y], and theirdifference, V[V-Y]. Implied 95% confidence intervals are given by dashed lines. 1 0.5 Correlation 0 V and Y V and T Y and T −0.5 5 10 15 20 Forecast horizon, daysFigure 6: Implied correlations between volatility (V), central tendency (Y), and their difference(T). Implied 95% confidence intervals are given by dashed lines. 32
  33. 33. V Y 0.8 V−Y 0.6 Autocorrelation 0.4 0.2 0 −0.2 20 40 60 80 Lag, daysFigure 7: Implied autocorrelations of volatility (V), central tendency (Y), and their difference(T) integrated over one day. Implied 95% confidence intervals are given by dashed lines. V Y 0.8 V−Y 0.6 Autocorrelation 0.4 0.2 0 −0.2 20 40 60 80 Lag, daysFigure 8: Implied autocorrelations of volatility (V), central tendency (Y), and their difference(T) integrated over 22 business days. Implied 95% confidence intervals are given by dashedlines. 33